Box CEO Aaron Levie posted on X on April 20 that rapid AI model improvement is forcing engineering teams to discard and rebuild their agent architectures every few quarters.
“It’s remarkable how often you need to be dramatically upgrading your AI architecture given the pace of progress in AI models right now,” Levie wrote. “If you’re building agents, you basically need to throw away large parts of previous work that you setup to compensate for model limitations every few quarters.”
He pointed to a specific category of technical debt: workarounds built for earlier model constraints. “The systems you built to mitigate context window limits aren’t useful anymore,” he wrote, noting that newer models can handle larger workloads with fewer patches. In some cases, teams can now “just throw more compute at a problem” in ways that were previously impractical, according to Benzinga’s coverage.
Enterprise Deployment Practices Shifting
Levie also flagged the velocity of change in enterprise deployment. “The way you would deploy agents in an enterprise 18 months ago is entirely different from the best practices that you’d have today,” he wrote.
The timing is notable. His post arrived the same weekend as a separate X thread noting that “most of tooling around LLMs was built for a world that largely doesn’t exist anymore,” listing RAG, GraphRAG, Multi Agent Orchestration, ReAct frameworks, prompt management tools, LLMOps tooling, eval tools, gateways, and finetuning libraries as categories that have been “obsoleted in the last 3 months.”
Levie’s Consistent Position
This is not Levie’s first public statement on agent velocity. In March, he told Business Insider that “much of the internet either needs to get rebuilt, or the software that we use will have to adapt. All software is going to have to be built for agents. That’s going to be a tectonic shift.”
Earlier this month, Levie said on Benzinga that AI would shift rather than eliminate jobs, noting that AI-driven efficiency creates new constraints rather than removing work.
The Rebuild Cost Problem
Levie’s comments surface a tension that enterprise teams building agents face now: the investment required to build agent infrastructure may depreciate faster than traditional software. If architectures need significant rework every few quarters, the total cost of ownership calculation for agent deployments shifts from “build once, maintain” to “build, discard, rebuild.”
For teams evaluating agent platforms, the practical question is whether to optimize for the current model generation or build abstraction layers that survive model upgrades. Levie’s own company, Box, has positioned itself on the infrastructure side of that trade, offering Box AI and Box Agent as layers that handle model transitions on behalf of enterprise customers.